Skip to content

Evaluate chatgpt's performance on the Open Information Extraction task using CaRB dataset

Notifications You must be signed in to change notification settings

PoloWitty/chatgpt-on-openie

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

chatgpt on OpenIE

This repo evaluate chatgpt open information extraction performance using CaRB (a benchmark on OpenIE task).

Poor result

chatgpt performs pretty poor on this task. It only achieve f1 of 0.248. On paperwithcode, the lowest score on this leaderboard is ~0.5 of f1 score. See more details here.

Chatgpt settings

api parameters

model: gpt-3.5-turbo
messages.role: user

other settings are all the same as default settings.

prompt

I input the following sentence to chatgpt on the brower, and choose one from its output.

provide five concise prompts or templates that can make you deal with the open information extraction task and output the orig_sentence, head, relation, tail, output them in one line and use [TAB] to seperate them

So the input prompt to api is actually as follows, the input sentences is attach to it.

"Prompt: Extract the relations between entities mentioned in the given sentence using open information extraction and output them in one line separated by tabs with head entity first. Output template: [ORIG_SENTENCE]\t[HEAD_ENTITY]\t[RELATION]\t[TAIL_ENTITY]\n"

Post process

There are a lot of format problems in the raw output. So I do some post processing. The details are as follows.

  • pick '\t' as '\t'
  • replace 'TAIL_ENTITY' by ''
  • replace 'HEAD_ENTITY' by ''
  • replace 'RELATION'/'Relation' by ''
  • replace '|' by '\t'
  • strip '(' and ')' for h,r,t

About

Evaluate chatgpt's performance on the Open Information Extraction task using CaRB dataset

Resources

Stars

Watchers

Forks